HAL Id: inria-00422651
https://hal.inria.fr/inria-00422651
Submitted on 8 Oct 2009
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Kget+: An efficient tool for managing VM image snapshots
Jérôme Gallard
To cite this version:
Jérôme Gallard. Kget+: An efficient tool for managing VM image snapshots. Toulouse’2009 (Ren-
Par’19 / SympA’13 / CFSE’7) – Poster Session, Sep 2009, Toulouse, France. �inria-00422651�
RenPar’19 / SympA’13 / CFSE’7 Toulouse, France, du 9 au 11 septembre 2009
Kget+ : An efficient tool for managing VM image snapshots
Jérôme Gallard
∗IRISA, INRIA – Centre Rennes Bretagne Atlantique, Rennes, France.
[email protected]
Résumé
Virtualization technologies have recently gained a lot of interest in Grid computing as they allow flexi- ble resource management. However, designing and implementing a complete framework to build VM images, to deploy and configure them automatically and to manage them (with migration and snapshot- ting operations) remains a real challenge. In this paper, after enumerating these challenges, we focus on the VM image snapshot management issue and present our tool, called Kget+, able to manage VM image snapshots in a very efficient way.
Mots-clés : Virtualization, Grid Resource Management, Efficient VM Image Copy, Kget+.
1. Introduction
Virtualization technologies provide flexible and powerful execution environments, offering isolation and security mechanisms, customization and encapsulation of entire application environments. More- over, thanks to their suspend/resume and migration capabilities, they allow the bare hardware to be strongly decoupled from the system software, which is a predominant feature in the context of dis- tributed architectures such as grids, which are exploited concurrently by multiple users. In this particu- lar context, VMs can be exploited to encapsulate jobs and then make grid resource management easier.
The challenge of managing applications on grids is moved to the problem of managing VMs on grids.
The rest of this paper is organized as follows : Section 2 introduces issues implied by the use of VM technologies in grids and presents more precisely the problem of VM snapshot management. In Section 3 we present Kget+, a tool able to manage VM image snapshots in an efficient way. Section 4 concludes and gives some perspectives of work.
2. Issues implied by the use of VM technologies in grids
Designing and implementing a framework able to manage user jobs in VMs at grid level requires to consider each step of a common job submission in a grid context : (i) VM image creation, (ii) VM image management (iii) job submission, VM deployment and job starting, (iv) VM/job life cycle (VM migra- tion, suspend/resume, snapshots, . . . ) and finally (v) job completion and VM shutdown.
In this document, we focus on the VM storage management issue and more precisely, on the VM image snapshot issue by proposing a tool, called Kget+, able to copy efficiently VM image snapshots from n nodes to one frontend (more information about the other issues could be found in [1]).
The VM image snapshot management issue
One drawback of the use of VMs in grids and clusters relates to the VM image snapshot management.
There are several methods to manage VM images (on each local node, on a remote server, ...), each method having their own good and bad points. In this paper we focus on the remote server method at cluster level by considering the copy of n VM image snapshots from n physical nodes to a frontend which implies a network bottleneck at the frontend side. To solve this problem, we have designed and implemented Kget+, which is presented in the next Section.
∗The INRIA team carries out this research work in the framework of the XtreemOS project partially funded by the European Commission under contract #FP6-033576.
3. Kget+ : Design and Experiments
Kget+ is our prototype allowing to copy VM image snapshots from n nodes to one frontend in an efficient way. It follows two principles : (i) to regulate the network bandwidth from the frontend point of view, (ii) to free a large part of the set of nodes in a minimum of time.
To do this, Kget+ makes sure that at all times one and only one node sends a VM image to the frontend (principle (i)). Parallelly, other nodes coordinate among themselves to group all VM images on a smaller number of nodes (principle (ii)). In this way, the network between all the nodes is used parallelly and the result is that several nodes are freed very quickly and the frontend network is never saturated. Of course this solution assumes that there is enough available disk space on the nodes to make the copies.
00:00:00 00:07:12 00:14:24 00:21:36 00:28:48 00:36:00 0
20 40 60 80 100 120
Kget+
Kaget scp
Time 00:01:55
Percentage of node freed
80
F
IG. 1 – Comparison between multiple scp , kaget and Kget+ . Kget+ allows to freed 80% of the nodes in less than 2 minutes.
To do this, we propose the following algorithm : While there is a node waiting to transfer a VM snap- shot to the frontend :
– copy one VM snapshot from the node containing the lowest number of VM snapshots
2to the frontend.
– from each other node, copy VM snapshots to another node except if :
– the destination node is already sending or receiv- ing a VM snapshot,
– the destination node has less VM snapshots than the sender node,
– the sender node is already sending or receiving a VM snapshot.
After implementing this algorithm, we made some evaluations of it on the Paravent cluster
from Rennes Grid’5000 site
3[2]. In our experiments we launch a job composed of 64 VM images on 64 physical nodes and we measure the time needed to copy the VM image snapshots from all nodes to the frontend with multiple simultaneous scp, with Kaget
4and with our tool Kget+. The size of one VM image snapshot is approximately 512MBytes. Figure 1 presents the percentage of nodes freed (Y- axis) according to the elapsed time (X-axis). This figure shows that, thanks to Kget+, a good scheduling of the copies gives very efficient results. Indeed, with Kget+ all nodes are freed in less than 15 minutes whereas with multiple scp and Kaget all nodes are freed in less than 18 and 34 minutes respectively.
In addition, with Kget+ 80% of the nodes are freed in less than 2 minutes (6% of the wall time).
4. Conclusion and Future Work
In this paper we present the design of Kget+, our prototype able to copy VM image snapshots from n physical nodes to one frontend in an efficient way. In addition we give results of preliminary ex- periments. This tool is able to schedule the copy of all the VM image snapshots from all the nodes to one frontend with the goal of : (i) regulating the network bandwidth from the frontend point of view, (ii) freeing a large part of the set of nodes in a minimum of time. This tool should be integrated with VMdeploy
5our VM management prototype able to deploy and manage VMs at grid level. This one needs improvements, however, it is a good framework to study one by one, all the issues implied by the use of VMs in grids which, from our point of view, will pave the way to scientific clouds.
Bibliographie
1. Jérôme Gallard et al. VMdeploy, Improving Best-Effort Job Management in Grid5000. Technical Report RR-6764, INRIA, December 2008.
2. Raphaël Bolze et al. Grid’5000 : A Large Scale and Highly Reconfigurable Experimental Grid Testbed. International Journal of High Performance Computing Applications, 20(4) :481–494, 2006.
2This could be easily extended to consider the whole size of VMs instead of the number.
3https ://www.grid5000.fr
4http ://taktuk.gforge.inria.fr/kanif/
5https ://www.grid5000.fr/mediawiki/index.php/VMdeploy
